LUMIC

A morphological profiling screen answers what a compound does to a cell by staining it and photographing it, but only for the combinations someone physically plated, and that grid grows faster than any robot can fill it. Generative priors could pick which wells are worth running, yet earlier Cell Painting generators learn each compound as a class label and paint one cell line, so a molecule outside the training panel has nothing to condition on and a new cell type has no way in.

LUMIC turns both conditioning signals into descriptions available for something never observed. A compound enters as its SMILES string, encoded by a hierarchical molecular graph autoencoder; the cell line enters as a self-supervised vision embedding of that line's own untreated control well. Generation happens first in embedding space — a one-dimensional diffusion model predicts the DINO embedding of the perturbed cell from the control embedding plus the chemical latent — and only then decoded into pixels by a cascade of two further diffusion models.

It comes from the Welch lab at the University of Michigan, with a co-author at MIT CSAIL.

#Key Features

• Compounds enter as structures, not labels: HGraph2Graph, trained on roughly 250,000 ZINC molecules plus the assay's own chemicals, supplies a 128-dimensional latent for any SMILES string, zero-padded to 256 dimensions.
• Cell identity comes from a control image: A DINO vision transformer embeds a 256 x 256 crop of the matched untreated well, so a new cell line is introduced by handing the model its control images rather than a class.
• Latent-first cascade: Predicting the perturbed embedding, decoding it to low resolution, and upsampling are three separately trained diffusion models, keeping sample quality high and the embedding usable on its own.
• Embeddings are a usable output: Classifiers trained on real embeddings recover cell type and the compound's target gene from generated ones, so analyses needing signal rather than pixels can stop after the first stage.

#Technical Details

Three DDPMs are trained independently on U-Net backbones adapted from the open imagen and denoising-diffusion-pytorch implementations. The first uses one-dimensional convolutions to diffuse over the 384-dimensional DINO embedding (75,000 steps); the second decodes an embedding into a 64 x 64 image (150,000 steps); the third takes the embedding and that low-resolution image to 256 x 256 (1,000,000 steps over ten days on three A40 GPUs). All use Adam, weight averaging, a 1,000-step cosine noise schedule and classifier-free guidance, sampled with DDIM at 250 steps.

Training combined 27,702 images from the JUMP Pilot Target-1 subset — 306 chemicals on U2OS and A549 cells at the 24-hour timepoint — with a new 3,168-image set covering 3T3, A549, HEK293T, HeLa and RPTE cells under a shared 61-compound panel. Three channels were stacked into RGB: Hoechst for DNA, phalloidin for actin, Golgi and plasma membrane, and MitoTracker for mitochondria. Thirty JUMP compounds were withheld, sixteen targeting genes never seen in training, along with ten compounds from the new set and every HeLa treatment image. On unconditional U2OS generation LUMIC reached a Kernel Inception Distance of 0.0154, against 0.1727 for IMPA and 0.3313 for PhenDiff. For unseen compounds on seen cell lines, generated images sat closer to real treated images than untreated controls did — KID 0.0512 versus 0.1133 on A549 — and an MLP trained on real DINO embeddings called cell type from generated embeddings at 0.918 accuracy, against 0.929 on real.

#Applications

The intended use is experiment design: generate the expected morphology for a candidate compound on a candidate cell line, then plate only the combinations whose predicted phenotype is informative. That extends a screen along both axes — new chemistry against profiled lines, profiled chemistry against a line only ever imaged untreated — suiting repurposing panels, toxicity triage, and synthetic image sets for methods development.

#Impact

The transferable idea is treating a perturbation as a style transferred onto a cell, the cell supplied as an image and the chemistry as a graph, so both sides of the grid stay open — unlike compound-conditioned generators such as MorphoDiff and the confounder-aware Cell Painting latent diffusion model, which condition on chemistry but not on a cell line's appearance. The authors state the limits plainly. Quality degrades for the fully held-out cell line, where within-class KID approaches baseline values, and CellProfiler features from generated images diverge from real ones: a cell-type classifier at 0.735 accuracy on held-out real features falls to 0.472 on generated ones, mostly on intensity-derived measurements. The corpus is modest — roughly 30,000 images, three dyes, a handful of cell lines — and the DINO encoder is fit to it rather than being a general microscopy backbone. Code and all three stages of weights are public, but the repository carries no license and the paper remains a preprint.